Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/PEOPLE/ZHAO CHENGYANG
// PERSON

Zhao Chengyang

ROLE CORE ENGINEER / FOUNDERAT RADIXARKMENTIONS 10LAST SEEN APRIL 30, 2026
// BIO

Zhao Chengyang (also known online as Chayenne Zhao) is a member of technical staff at RadixArk, a San Francisco-based AI infrastructure company launched in May 2026 that develops and commercializes SGLang. He is one of the principal maintainers of SGLang, an open-source, high-performance serving framework for large language models, and focuses on optimizing reinforcement learning training pipelines and large-scale inference systems. He studied computer science at Tsinghua University before pursuing a Ph.D. at UCLA and is recognized for his technical contributions to end-to-end RL infrastructure and multi-turn agent inference at frontier scale.

// RECENT MENTIONS
// SIGNALS
10 SIGNALS
01
product·晚点聊 LateTalk·APRIL 30, 2026

SGLang is deployed on over 400,000 GPUs globally for production-level inference — one of the largest open-source engines of this generation.

Source
02
product·晚点聊 LateTalk·APRIL 30, 2026

TileLang is now being used by frontier labs as one of the default choices for implementing algorithms — this has happened in roughly the last year and a half.

Source
03
product·晚点聊 LateTalk·APRIL 30, 2026

Claude's generation of models, from 4.5 onwards, in multi-step agentic code performance compared to before, has truly improved enormously.

Source
04
mention·晚点聊 LateTalk·APRIL 30, 2026

In inference, the gap between proprietary and open-source is not that large. But in training, proprietary training still leads open-source by quite a bit. A model might launch in February, but it might take until May or June before an open-source RL framework can actually run its RL pipeline.

Source
05
mention·晚点聊 LateTalk·APRIL 30, 2026

Our team did substantial engineering optimization and successfully ran both the inference and RL pipelines on the day DeepSeek V4 was released.

Source
06
mention·晚点聊 LateTalk·APRIL 30, 2026

TileLang has now been adopted by frontier labs as one of the default choices for algorithm implementation.

Source
07
mention·晚点聊 LateTalk·APRIL 30, 2026

If 3.0 gets into WeChat, the competitive landscape could get very interesting.

Source
08
mention·晚点聊 LateTalk·APRIL 30, 2026

TileLang is now being used by cutting-edge labs as one of their primary choices for algorithm implementation.

Source
09
mention·晚点聊 LateTalk·APRIL 30, 2026

Our team made substantial engineering optimizations and successfully ran both the inference and RL pipelines on the day V4 was released.

Source
10
mention·晚点聊 LateTalk·APRIL 30, 2026

For example, MTP is quite critical for voice — like when you open Doubao and talk to it, the speed at which it produces the first piece of audio is very fast. Very unfortunately, on the open-source side we haven't done this nearly as well.

Source

AI-extracted from podcast / newsletter / paper summaries. May contain errors.